Maya: Hey builders! Here's what's on the radar today on Claude Code Review. Claude Managed Agents hit public beta, Claude Code can now coordinate up to a thousand agents, cloud sessions keep jobs running after you close the laptop, version two point one point two nine six lands, and Max plans add API credits. I'm Maya.

James: And I'm James. Let's get into it.

Maya: Starting with Claude Managed Agents, which entered public beta on October eighth, according to the ClaudeDevs community account. The pitch is dynamic workflows: one lead agent drafts a plan, splits it into phases, hands each phase to a separate agent, then stitches the results together automatically.

James: That's a clean mechanism, Maya. The linked blog post shows a concrete setup: an agent that reads Slack and GitHub on a schedule, keeps credentials and memory between runs, and posts status updates back to the team. You spin it up with a Claude Code command.

Maya: Exactly right, and it draws on new API credits bundled into the Max and Team plans. ClaudeDevs frames it for your most ambitious workloads, so long-running, multi-step jobs, not quick one-off prompts. It's still public beta, so plenty is unspecified.

James: That's the natural bridge to story two, Maya. An October ninth post from ClaudeDevs added a configuration letting a single Claude Code run spread work across up to one thousand agents at once, with Claude itself handling the orchestration.

Maya: Great catch on the connection, James. To turn it on, you set an agent's multiagent type to multiagent two zero two six one zero zero one, then just ask Claude to run a workflow. Claude writes the plan and orchestrates however many agents the job needs.

James: And the tradeoff is cost. Anthropic's post says these dynamic workflows are powerful and can use a lot of tokens, so they recommend starting with a scoped, narrow task and scaling up only once you see how the orchestration behaves.

Maya: Exactly, and to try it you run slash claude-api managed-agents-onboard bug-hunter for a working example, or pull a template from the agent-quickstart page at platform dot claude dot com.

James: Story three is cloud sessions, from an Anthropic walkthrough on October thirteenth. Claude clones your repo from GitHub into a virtual machine and works on its own branch. Close the laptop and the job keeps running; Claude can even ask a question you answer from your phone.

Maya: The use case is work that outlasts your time at the desk, James, a big refactor or a migration you hand off at the end of the day. Setup is two steps: connect your GitHub account, then install the Claude GitHub App and choose which repos it can see.

James: And here's the catch worth flagging: each session starts from a fresh clone. Committed and pushed work is there; anything only on your machine, personal settings, skills, plugins, is not. For work depending on local config, Anthropic says run Claude Code locally instead.

Maya: Fair flag, James. Sessions count toward your plan's limits like local ones, with no extra charge for the virtual machine. Review the diff on your phone, tap Create PR, and Team and Enterprise plans can run cloud sessions on their own infrastructure.

James: Story four is Claude Code version two point one point two nine six, shipped October ninth. The headline change: a new autoCompactWindow field lets a subagent summarize its own history to free up context earlier than the main conversation does.

Maya: That's the one builders will feel, James. The cost figures also updated: slash cost, the status line, and budget checks now price Sonnet five point five cache reads at ten cents per million tokens, down from twenty cents.

James: And a few fixes worth a look, Maya. The default cap on MCP tool descriptions doubles from two thousand forty-eight to four thousand ninety-six characters, and the Read tool gains an allow_large option to pull a whole oversize file in one call when context has room.

Maya: Last story: Max and Team plans now include monthly API credits, announced October eighth by ClaudeDevs. One hundred dollars on Max five X, two hundred on Max twenty X, and up to five hundred pooled across a Team plan, working with any Claude model.

James: And the limit that matters most, Maya: those credits cannot be used in interactive Claude Code sessions. They cover the Messages API, Managed Agents, the Agent SDK, and third-party harnesses that accept a Claude API key.

Maya: Right, and ClaudeDevs calls Managed Agents the quickest way to use them. To claim, go to Settings, Billing on claude dot ai, link a Console organization, and new subscribers can claim after seven days.

Maya: That's your Claude Code Review daily brief. Check claudecodereview.ai for the full stories. Keep building.